Download ppt - TMVA - SourceForge

HEP Data Mining with TMVA
 ToolKit for Multivariate Analysis with ROOT 
Andreas Hoecker(*) (CERN)
Seminar, IFJ – Krakow, Feb 27, 2007
(*)
Krakow Seminar, Feb 27, 2007
on behalf of J. Stelzer, F. Tegenfeldt, H.+K. Voss, and many other contributors
A. Hoecker: Data Mining with TMVA
1
advert isement
We (finally) have a
Users Guide !
Please check on tmva.sf.net
for its imminent release
TMVA Users Guide
68pp, incl. code examples
to be submitted to arXiv:physics
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
2
Preliminary Remarks
One example for “Machine Learning”:
It is not so useful to let the machine learn what
we know well, but let it do where we’re bad:
Human
Set the goal (define “signal” and “background”)
Determine discriminating variables
Find appropriate control samples
Machine
Develop machine learning algorithms
Train the algorithm
Independently evaluate the algorithm
Apply the algorithm to unknown data
It is good to understand what the machine does – but it is not mandatory !
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
3
Event Classification
Suppose data sample with two types of events: H0, H1
We have found discriminating input variables x1, x2, …
What decision boundary should we use to select events of type H1 ?
Rectangular cuts?
x2
A linear boundary?
H1
H0
x2
H1
H0
x1
A nonlinear one?
x2
H1
H0
x1
x1
How can we decide this in an optimal way ?  Let the machine learn it !
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
4
Multivariate Event Classification
All multivariate classifiers have in common to condense (correlated)
multi-variable input information in a single scalar output variable
It is a RnR regression problem; classification is in fact a discretised regression
y(H0)  0, y(H1)  1
…
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
5
Event Classification in High-Energy Physics (HEP)
Most HEP analyses require discrimination of signal from background:
Event level (Higgs searches, …)
Cone level (Tau-vs-jet reconstruction, …)
Track level (particle identification, …)
Lifetime and flavour tagging (b-tagging, …)
Parameter estimation (CP violation in B system, …)
etc.
The multivariate input information used for this has various sources
Kinematic variables (masses, momenta, decay angles, …)
Event properties (jet/lepton multiplicity, sum of charges, …)
Event shape (sphericity, Fox-Wolfram moments, …)
Detector response (silicon hits, dE/dx, Cherenkov angle, shower profiles, muon hits, …)
etc.
Traditionally few powerful input variables were combined; new methods
allow to use up to 100 and more variables w/o loss of classification power
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
6
Multivariate Classification Algorithms
Traditional
Rectangular cuts (optimisation often “by hand”)
Variants
A large variety of multivariate classifiers (MVAs) exists
Prior decorrelation of input variables (input to cuts and likelihood)
Projective likelihood (up to 2D)
Linear discriminant analysis (2 estimators, Fisher)
Nonlinear discriminant analysis (Neural nets)
Principal component analysis of input variables
Multidimensional likelihood (kernel nearest neighbor methods)
New
Decision trees with boosting and bagging, Random forests
Rule-based learning machines
Support vector machines
Bayesian neural nets, and more general Committee classifiers
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
7
Multivariate Classification Algorithms
How to dissipate (often diffuse) skepticism against the use of MVAs
black boxes !
Certainly, cuts are transparent, so
• if cuts are competitive (rarely the case)  use them
• in presence of correlations, cuts loose transparency
• “we should stop calling MVAs black boxes and understand how they behave”
what if the
training samples
incorrectly describe the data ?
Not good, but not necessarily a huge problem:
how can
one evaluate
systematics ?
There is no principle difference in systematics evaluation between
single discriminating
variables and MVAs
variables and MVA
Krakow Seminar, Feb 27, 2007
•
•
•
•
performance on real data will be worse than training results
however: bad training does not create a bias !
only if the training efficiencies are used in data analysis  bias
optimized cuts are not in general less vulnerable to systematics (on the contrary !)
• need control sample for MVA output (not necessarily for each input variable)
A. Hoecker: Data Mining with TMVA
8
TMVA
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
9
What is TMVA
The various classifiers have very different properties
Ideally, all should be tested for a given problem
Systematically choose the best performing classifier
Comparisons between classifiers improves the understanding and takes away mysticism
TMVA ― Toolkit for multivariate data analysis with ROOT
Framework for parallel training, testing, evaluation and application of MV classifiers
A large number of linear, nonlinear, likelihood and rule-based classifiers implemented
Each classifier provides a ranking of the input variables
The input variables can be decorrelated or projected upon their principal components
Training results and full configuration are written to weight files and applied by a Reader
Clear and simple user interface
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
10
TMVA Development and Distribution
TMVA is a sourceforge (SF) package for world-wide access
Home page ………………. http://tmva.sf.net/
SF project page …………. http://sf.net/projects/tmva
View CVS …………………http://tmva.cvs.sf.net/tmva/TMVA/
Mailing list .………………..http://sf.net/mail/?group_id=152074
ROOT class index ………. http://root.cern.ch/root/htmldoc/TMVA_Index.html
Very active project  fast response time on feature requests
Currently 4 main developers, and 24 registered contributors at SF
~ 700 downloads since March 2006 (not accounting cvs checkouts and ROOT users)
Written in C++, relying on core ROOT functionality
Full examples distributed with TMVA, including analysis macros and GUI
Scripts are provided for TMVA use in ROOT macro, as C++ executable or with python
Integrated and distributed with ROOT since ROOT v5.11-03
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
11
The TMVA Classifiers
Currently implemented classifiers :
Rectangular cut optimisation
Projective and multidimensional likelihood estimator
Fisher and H-Matrix discriminants
Artificial neural networks (three different multilayer perceptrons)
Boosted/bagged decision trees with automatic node pruning
RuleFit
In work :
Support vector machine
Committee classifier
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
12
Data Preprocessing: Decorrelation
Commonly realised for all methods in TMVA (centrally in DataSet class)
Removal of linear correlations by rotating input variables
Determine square-root C of correlation matrix C, i.e., C = CC
Compute C by diagonalising C: D  ST CS  C  S DST
Transform original (x) into decorrelated variable space (x) by: x = C1x
Various ways to choose basis for decorrelation (also implemented PCA)
Note that decorrelation is only complete, if
Correlations are linear
Input variables are Gaussian distributed
Not very accurate conjecture in general
original
Krakow Seminar, Feb 27, 2007
SQRT derorr.
A. Hoecker: Data Mining with TMVA
PCA derorr.
13
Rectangular Cut Optimisation
Simplest method: cut in rectangular variable volume
xcut  i event   0,1 
vvariables
 x  i    x
v
event
v ,min
, xv ,max 

Usually training files in TMVA do not contain realistic signal and
background abundance
Cannot optimize for best significance
Instead: scan in signal efficiency [01] and maximise background rejection
From this scan, the best working point (cuts) for any sig/bkg numbers can be derived
Technical challenge: how to find optimal cuts
MINUIT fails due to non-unique solution space
TMVA uses: Monte Carlo sampling, Genetics Algorithm, Simulated Annealing
Huge speed improvement of volume search by sorting events in binary tree
Cuts usually benefit from prior decorrelation of cut variables
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
14
Projective Likelihood Estimator (PDE Approach)
Much liked in HEP: probability density estimators for
each input variable combined in likelihood estimator
Likelihood ratio
for event ievent
y L  i event  
PDFs

kvariables
discriminating variables
pksignal  xk (i event )
PDE introduces fuzzy logic


U
  pk  xk (i event ) 

Uspecies  kvariables

Species: signal,
background types
Ignores correlations between input variables
Optimal approach if correlations are zero (or linear  decorrelation)
Otherwise: significant performance loss
yL often strongly peaked 0,1: transform output (configurable)


y L  y L   1 ln y L1  1
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
15
PDE Approach: Estimating PDF Kernels
Technical challenge: how to estimate the PDF shapes
3 ways:
parametric fitting (function)
Difficult to automate
for arbitrary PDFs
nonparametric fitting
Easy to automate, can create
artefacts/suppress information
event counting
Automatic, unbiased,
but suboptimal
We have chosen to implement
nonparametric fitting in TMVA
Binned shape interpolation using spline
functions (orders: 1, 2, 3, 5)
original
distribution
is Gaussian
Unbinned kernel density estimation
(KDE) with Gaussian smearing
TMVA performs automatic validation of
goodness-of-fit
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
16
Multidimensional PDE Approach
Use a single PDF per event class (sig, bkg), which spans Nvar dimensions
PDE Range-Search: count number of signal and background events in
“vicinity” of test event  preset or adaptive volume defines “vicinity”
yPDERS  ievent ,V 
The signal estimator is then given by (simplified,
full formula accounts for event weights and training population)
PDE-RS ratio
for event ievent
chosen
volume
yPDERS  i event ,V  
Carli-Koblitz, NIM
A501, 576 (2003)
x2
0.86
H1
#signal events in V
nS  i event ,V 
nS  i event ,V   nB  i event ,V 
test
event
H0
#background events in V
x1
Improve yPDERS estimate within V by using various Nvar-D kernel estimators
Enhance speed of event counting in volume by binary tree search
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
17
Fisher Linear Discriminant Analysis (LDA)
Well known, simple and elegant classifier
LDA determines axis in the input variable hyperspace such that a
projection of events onto this axis pushes signal and background
as far away from each other as possible
Classifier computation couldn’t be simpler:
yFi  i event   F0 

kvariables
“Fisher coefficients”
xk  i event   Fk
Nvar
Fisher coefficients given by: Fk  Wk1  xS,  xB,  , where W is sum CS + CB
1
Fisher requires distinct sample means between signal and background
Optimal classifier for linearly correlated Gaussian-distributed variables
H-Matrix: poor man’s version of Fisher discriminant
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
18
Nonlinear Analysis: Artificial Neural Networks
Achieve nonlinear classifier response by “activating”
output nodes using nonlinear weights
Feed-forward Multilayer Perceptron
Call nodes “neurons” and arrange them in series:
1 input layer
1
..
.
Nvar discriminating
input variables
i
..
.
N
xi(0)
1..Nvar
w11
w1j
w ij
k hidden layers
...
1
..
.
1 ouput layer
1
..
.
j
..
.
Mk
2 output classes
(signal and background)
( k 1)
x1,2
M1
x(j k )
Mk 1


 A  w0( kj )   wij( k )  xi( k 1) 
i 1


(“Activation” function)
with:

A( x )  1  e  x

Weierstrass theorem: can
approximate any continuous
functions to arbitrary precision
with a single hidden layer and
an infinite number of neurons
1
Three different multilayer perceptrons available in TMVA
Adjust weights (=training) using “back-propagation”:
For each training event compare desired and received MLP outputs  {0,1}: ε = d – r
Correct weights, depending on ε and a “learning rate” η
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
19
Decision Trees
Sequential application of cuts splits the data into
nodes, where the final nodes (leafs) classify an
event as signal or background
Growing a decision tree:
Start with Root node
Split training sample according to
cut on best variable at this node
Splitting criterion: e.g., maximum
“Gini-index”: purity  (1– purity)
Continue splitting until min. number
of events or max. purity reached
Classify leaf node according to majority of events, or give
Decision
tree before
pruningtest events are classified accordingly
weight;
unknown
Decision tree
after pruning
Bottom-up pruning of a decision tree
Remove statistically insignificant nodes to reduce tree overtraining  automatic in TMVA
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
20
Boosted Decision Trees (BDT)
Data mining with decision trees is popular in science (so far mostly outside of HEP)
Advantages:
Easy interpretation – can always be represented in 2D tree
Independent of monotonous variable transformations, immune against outliers
Weak variables are ignored (and don’t (much) deteriorate performance)
Shortcomings:
Instability: small changes in training sample can dramatically alter the tree structure
Sensitivity to overtraining ( requires pruning)
Boosted decision trees: combine forest of decision trees, with differently
weighted events in each tree (trees can also be weighted), by majority vote
e.g., “AdaBoost”: incorrectly classified events receive larger weight in next decision tree
“Bagging” (instead of boosting): random event weights, resampling with replacement
Boosting or bagging are means to create set of “basis functions”: the final classifier is
linear combination (expansion) of these functions
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
21
Predictive Learning via Rule Ensembles (RuleFit)
Following RuleFit approach by Friedman-Popescu
Friedman-Popescu, Tech Rep,
Stat. Dpt, Stanford U., 2003
Model is linear combination of rules, where a rule is a sequence of cuts
RuleFit classifier
rules (cut sequence
 rm=1 if all cuts
satisfied, =0 otherwise)
MR

normalised
discriminating
event variables
nR
yRF  x   a0   am rm xˆ   bk xˆk
m 1
Sum of rules
k 1
Linear Fisher term
The problem to solve is
Create rule ensemble: use forest of decision trees
Fit coefficients am, bk: “gradient direct regularization” (Friedman et al.)
Fast, rather robust and good performance
One of the elementary cellular automaton rules (Wolfram 1983, 2002). It specifies the next color in a cell, depending
on its color and its immediate neighbors. Its rule outcomes are encoded in the binary representation 30=000111102.
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
22
Using TMVA
A typical TMVA analysis consists of two main steps:
1.
Training phase: training, testing and evaluation of classifiers using data samples with
known signal and background composition
2.
Application phase: using selected trained classifiers to classify unknown data samples
Illustration of these steps with toy data samples
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
23
Code Flow for Training and Application Phases
Can be ROOT scripts, C++ executables or python scripts (via PyROOT),
or any other high-level language that interfaces with ROOT
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
24
A Toy Example (idealized)
Use data set with 4 linearly correlated Gaussian distributed variables:
--------------------------------------Rank : Variable : Separation
--------------------------------------1 : var3
: 3.834e+02
2 : var2
: 3.062e+02
3 : var1
: 1.097e+02
4 : var0
: 5.818e+01
---------------------------------------
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
25
Preprocessing the Input Variables
Decorrelation of variables before training is useful for this example
Similar distributions for PCA
Note that in cases with non-Gaussian distributions and/or nonlinear correlations
decorrelation may do more harm than any good
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
26
Validating the Classifier Training
Projective likelihood PDFs, MLP training, BDTs, …
TMVA GUI
average no. of nodes before/after pruning: 4193 / 968
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
27
Testing the Classifiers
Classifier output distributions for independent test sample:
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
28
Evaluating the Classifiers
There is no unique way to express the performance of a classifier
 several benchmark quantities computed by TMVA
Signal eff. at various background effs. (= 1 – rejection) when cutting on classifier output
1  yˆS ( y )  yˆ B ( y )
dy
2  yˆS ( y )  yˆ B ( y )
2
The Separation:
The discrimination Significance:  y S  y B 
 y2,S   y2,B
The average of the signal -transform:  y   yˆS ( y ) dy
(the -transform of a classifier yields a uniform background distribution,
so that the signal shapes can be directly compared among the classifiers)
Remark on overtraining
Occurs when classifier training has too few degrees of freedom because the classifier
has too many adjustable parameters for too few training events
Sensitivity to overtraining depends on classifier: e.g., Fisher weak, BDT strong
Compare performance between training and test sample to detect overtraining
Actively counteract overtraining: e.g., smooth likelihood PDFs, prune decision trees, …
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
29
Better classifier
Evaluating the Classifiers (taken from TMVA output…)
Check
for overtraining
Evaluation results ranked by best signal efficiency and purity (area)
-----------------------------------------------------------------------------MVA
Signal efficiency at bkg eff. (error): | SepaSignifiMethods:
@B=0.01
@B=0.10
@B=0.30
Area
| ration: cance:
-----------------------------------------------------------------------------Fisher
: 0.268(03) 0.653(03) 0.873(02) 0.882 | 0.444
1.189
MLP
: 0.266(03) 0.656(03) 0.873(02) 0.882 | 0.444
1.260
LikelihoodD
: 0.259(03) 0.649(03) 0.871(02) 0.880 | 0.441
1.251
PDERS
: 0.223(03) 0.628(03) 0.861(02) 0.870 | 0.417
1.192
RuleFit
: 0.196(03) 0.607(03) 0.845(02) 0.859 | 0.390
1.092
HMatrix
: 0.058(01) 0.622(03) 0.868(02) 0.855 | 0.410
1.093
BDT
: 0.154(02) 0.594(04) 0.838(03) 0.852 | 0.380
1.099
CutsGA
: 0.109(02) 1.000(00) 0.717(03) 0.784 | 0.000
0.000
Likelihood
: 0.086(02) 0.387(03) 0.677(03) 0.757 | 0.199
0.682
-----------------------------------------------------------------------------Testing efficiency compared to training efficiency (overtraining check)
-----------------------------------------------------------------------------MVA
Signal efficiency: from test sample (from traing sample)
Methods:
@B=0.01
@B=0.10
@B=0.30
-----------------------------------------------------------------------------Fisher
: 0.268 (0.275)
0.653 (0.658)
0.873 (0.873)
MLP
: 0.266 (0.278)
0.656 (0.658)
0.873 (0.873)
LikelihoodD
: 0.259 (0.273)
0.649 (0.657)
0.871 (0.872)
PDERS
: 0.223 (0.389)
0.628 (0.691)
0.861 (0.881)
RuleFit
: 0.196 (0.198)
0.607 (0.616)
0.845 (0.848)
HMatrix
: 0.058 (0.060)
0.622 (0.623)
0.868 (0.868)
BDT
: 0.154 (0.268)
0.594 (0.736)
0.838 (0.911)
CutsGA
: 0.109 (0.123)
1.000 (0.424)
0.717 (0.715)
Likelihood
: 0.086 (0.092)
0.387 (0.379)
0.677 (0.677)
-----------------------------------------------------------------------------
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
30
Evaluating the Classifiers (with a single plot…)
Smooth background rejection versus signal efficiency curve:
(from cut on classifier output)
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
31
Remarks
Additional information provided during the TMVA training phase
Classifiers provide specific ranking of input variables
Correlation matrix and classification “overlap” matrix for classifiers: if two classifiers have
similar performance, but significant non-overlapping classifications  combine them!
Such a combination of classifiers is denoted a “Committee classifier”: currently under
development (BDTs and RuleFit are committees of base learners, Bayesian ANNs are committees of ANNs, etc)
PDFs for the classifier outputs can be used to derive signal probabilities:
PMVA  i event  
fS yˆS (i event )
, with fS the signal fraction in sample
fS yˆS (i event )  1  fS  yˆ B (i event )
All classifiers write ROOT and text weight files for configuration and training results
 feeds the application phase (Reader)
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
32
More Toy Examples
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
33
More Toys: Linear-, Cross-, Circular Correlations
Illustrate the behaviour of linear and nonlinear classifiers
Linear correlations
Linear correlations
Circular correlations
(same for signal and background)
(opposite for signal and background)
(same for signal and background)
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
34
How does linear
decorrelation
affect strongly
nonlinear cases ?
Original correlations
SQRT decorrelation
PCA decorrelation
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
35
Weight Variables by Classifier Performance
How well do the classifier resolve the various correlation patterns ?
Linear correlations
Linear correlations
Circular correlations
(same for signal and background)
(opposite for signal and background)
(same for signal and background)
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
36
Final Classifier Performance
Background rejection versus signal efficiency curve:
Linear
Circular
Cross
Example
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
37
Stability with Respect to Irrelevant Variables
Toy example with 2 discriminating and 4 non-discriminating variables ?
use all
only
discriminant
two discriminant
variables in classifiers
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
38
Summary & Plans
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
39
Summary of the Classifiers and their Properties
Classifiers
Criteria
Performance
no / linear
correlations
nonlinear
correlations
Training
Speed
Response
Robustness
Overtraining
Weak input
variables
Course of
dimensionality
Clarity
Krakow Seminar, Feb 27, 2007
Rectangular
Cuts
Projective
Likelihood
Multi-D
PDERS
H-Matrix
Fisher
MLP
BDT
RuleFit
































































A. Hoecker: Data Mining with TMVA
40
Summary & Plans
TMVA unifies highly customizable and performing multivariate classification algorithms in a single user-friendly framework
This ensures most objective classifier comparisons, and simplifies their use
TMVA is available on tmva.sf.net, and in ROOT ( 5.11/03)
A typical TMVA analysis requires user interaction with a Factory (for the
classifier training) and a Reader (for the classifier application)
ROOT Macros are provided for the display of the evaluation results
Forthcoming:
Imminent: TMVA version 3.6.0 with new features (together with a detailed Users Guide)
Support Vector Machine (M. Wolter & A. Zemla)
Bayesian classifiers
Committee method
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
41
Copyrights & Credits
TMVA is open source !
Use & redistribution of source permitted according to terms in BSD license
Several similar data mining efforts with rising importance in most fields of
science and industry
Important for HEP:
Parallelised MVA training and evaluation pioneered by Cornelius package (BABAR)
Also frequently used: StatPatternRecognition package by I. Narsky
Many implementations of individual classifiers exist
Acknowledgments: The fast development of TMVA would not have been possible without the contribution and feedback from
many developers and users to whom we are indebted. We thank in particular the CERN Summer students Matt Jachowski (Stanford) for the implementation of TMVA's new MLP neural network, and Yair Mahalalel (Tel Aviv) for a significant improvement of
PDERS. We are grateful to Doug Applegate, Kregg Arms, Ren\'e Brun and the ROOT team, Tancredi Carli, Elzbieta Richter-Was,
Vincent Tisserand and Marcin Wolter for helpful conversations.
Krakow Seminar, Feb 27, 2007
A. Hoecker: Data Mining with TMVA
42